claude-context

AI coding agents are strong reasoners but weak retrievers. When a codebase grows past a few hundred files, an agent cannot hold all of it in context, so it falls back to grep, file globbing, or loading whole directories, which burns tokens and still misses code that does not match the exact search string. claude-context, maintained by Zilliz (the team behind the Milvus vector database), tackles this gap. It is an MCP server that indexes a repository once and then exposes a semantic code-search tool to whatever agent you connect, whether that is Claude Code, Cursor, Gemini CLI, Codex CLI, Windsurf, or others. The idea is simple: let the agent ask for the code that handles user authentication and get the relevant chunks back, rather than guessing at filenames.

The pipeline has a few distinctive pieces. Code is split with an AST-based chunker that respects function and class boundaries, with a fallback splitter when parsing fails, so the embedded units are coherent rather than arbitrary line windows. Each chunk is turned into a vector via a configurable embedding provider: OpenAI, VoyageAI (including the code-tuned voyage-code-3), Gemini, or a local Ollama model. Vectors land in Milvus or the managed Zilliz Cloud. At query time it runs hybrid search, combining BM25 keyword matching with dense-vector similarity, which catches both exact identifiers and conceptual matches. The most notable engineering choice is Merkle-tree incremental indexing: only files that actually changed get re-embedded, so keeping a large repo in sync stays cheap as the code evolves.

The clearest fit is large monorepos and sprawling codebases where an agent realistically cannot see everything and where grep-style search keeps returning either too much or the wrong thing. Teams running Claude Code or Cursor across millions of lines are the obvious audience, as are people building custom agents on LangChain or LangGraph who want a drop-in retrieval layer. Concrete uses include locating every place a payment flow is touched before a refactor, mapping how two distant modules relate, or giving a new agent session enough grounding to answer architecture questions without manually pasting files. It ships as npm packages, an MCP server, and a VS Code extension branded Semantic Code Search, so adoption does not force one specific workflow.

The honest caveats are real. claude-context needs an embedding-provider API key, which adds recurring cost and an external dependency on every index and many queries, unless you run Ollama locally and accept lower embedding quality. It also expects a vector database: Milvus to self-host or a Zilliz Cloud account, which is more setup than a pure local tool. The headline figure of roughly 40% token reduction is the project's own claim under its own equivalent-retrieval-quality condition, not an independent benchmark, so treat it as directional. There is real overlap with other code-search MCP servers emerging in this space, and the obvious incentive question: Zilliz benefits when you store vectors in its database, so the steering toward Milvus and Zilliz Cloud is unsurprising.

Against a commercial option like Sourcegraph Cody, claude-context is narrower and more composable. Cody is a polished, hosted code-intelligence product with its own UI, enterprise controls, and search backend; claude-context is a focused MCP building block that plugs retrieval into the agent you already use, with the embedding and storage choices left in your hands. That openness is the appeal for teams that want to own the stack and avoid lock-in, and the Zilliz backing lends credibility on the vector-infrastructure side. Adopt it when you have a genuinely large codebase, agents that already speak MCP, and tolerance for embedding costs and a vector store. Skip it when your project fits comfortably in context or when adding an external embedding API is not worth the overhead.